The present invention relates to signal processing and, more particularly, to signal processing systems for extracting communication signals from environments containing uncorrelated co-channel interference.
The problem of extracting a signal from a noisy environment is well known in the signal processing arts. The fundamental problem facing any receiver designer is how to improve the reception of a desired signal in the presence of unknown and undesired interfering signals, channel distortion, and thermal background noise. In principle, this can be accomplished by signal processing. For example, consider a multisensor receiver consisting of an array of spatially-separated antennae which is receiving a desired signal from a first direction and interfering signals from other directions. By forming the appropriate linear combination of the sensor outputs, the signals arriving from the desired direction will be accentuated while signals from other directions are attenuated. Similarly, in single sensor receivers, notch filters can be used to place notches at the frequencies of narrowband interfering signals and other filter types can be used to equalize linear channel distortion. In both cases, the desired signal reception can be significantly improved by passing the received signal (or signals) through a linear combiner with the proper combiner weights.
The basic problem is to set the combiner weights. If all of the parameters of the interfering signals are known, the proper combiner weights can be calculated. Unfortunately, such information is not always available. In many practical communications problems, the dominant type of channel corruption may be known. However, even in those cases in which the type of corruption is known, it may not be possible to know the numb type of corruption is known, it may not be possible to know the number, strength, or directions-of-arrival of the interfering signals at any given time. For example, in mobile radio applications, these interference parameters will depend on the time of day and the physical location of the transmitter and receiver. In this case, not only will the parameters be unknown at the beginning of the transmission, but also, they may vary during the transmission.
In such environments, it is usually impossible to preset the linear combiner weights to significantly improve reception of the desired signal. Hence, an adaptive algorithm must be used to learn the correct weight settings and to vary these settings over the course of the transmission if the parameters vary. This is accomplished by exploiting some known characteristic of the desired signal that distinguishes it from the interfering signals and noise.
Prior art methods for calculating the combiner weights have focused primarily on methods that optimize some measure of signal quality. A number of different quality measures have been used. For example, the Applebaum algorithm (S. P. Applebaum, "Adaptive Arrays," Syracuse University Research Corporation," Rep. SPLTR66-1, August 1966) maximizes the signal-to-noise ratio at the output to the array. The Widrow-hoff least-mean-square algorithm (B. Widrow, "Adaptive Filters I: Fundamentals" Stanford University Electronics Laboratory, Systems Theory Laboratory, Center for Systems Research, Rep. SU-SEL-66-12, Tech. Rep. 6764-6, December 1966) minimizes the mean-square-error between the desired signal and the output of an adaptive array.
More recently, exact least squares algorithms which optimize deterministic, time-averaged, measures of output signal quality have been developed (P. E. Mantey, L. J. Griffiths, "Iterative Least-Squares Algorithms for Signal Extraction," Second Hawaii Conference on Systems Science, Jan. 1969, pp. 767-770. B. Friedlander, "Lattice Filters for Adaptive Processing," Proc. IEEE, pp. 879-867, August 1982). In directly implementing an adaptive processor that optimizes any such quality measure, the receiver designer must have accurate knowledge of the cross-correlation between the transmitted and received signals. In practice, this requires close cooperation between the receiver and the desired-signal transmitter. In applications where the cross-correlations are not known at the start of the desired-signal transmission, these statistics must be learned by the receiver at the start of the transmission and, on tune-varying channels, updated over the course of the transmission.
The earliest methods for accomplishing this in telecommunications applications required the transmitter to send a known signal over the channel. This signal was sent at the beginning of the transmission or intermittently in lieu of the information-bearing signal. In this manner, the receiver could be trained at the start of the transmission and the combiner weights updated during the transmission. Other embodiments of this type of system transmit a pilot signal along with the information-bearing signal. This pilot signal is used to train and continuously adapt the receiver. These methods are effective in those situations in which there is cooperation between the transmitter and receiver.
In many applications, however, quality-optimizing techniques cannot be directly used. For example, the channel may be varying too rapidly for start-up or intermittent adaptation to be effective. In addition, the system resources (power, dynamic-range, bandwidth, etc.) may be too limited to allow pilot signals to be added to the information-bearing signal.
Alternatively, the receiver may not have the necessary control over the transmitter. This is the case when the receiver is a natural source such as a person speaking, or when the receiver that must be adapted is not the intended receiver in the communication channel, e.g., in reconnaissance applications.
In applications in which a known desired signal cannot be made available to the receiver, the designer must make use of blind adaptation techniques that exploit other observable properties of the desired signal or the environment in which the signal is transmitted. Prior art algorithms for accomplishing this may be divided into three classes. In the first class of techniques, referred to as demodulation-directed techniques, a reference signal is produced by demodulating and remodulating the processor output signal. This reference signal is then used as a training signal in a conventional adaptive processing algorithm. This technique is commonly employed in decision-directed and decision-feedback equalizers in telephony systems (J. G. Proakis, "Advances in Equalization of Intersymbol Interference," Advances in Communications Systems, ed. by A. V. Balakrishnan, A. J. Viterbi, N.Y. Academic Press, 1975). It has also been used to adapt antenna arrays in spread-spectrum communication systems (R. T. Compton, "An Adaptive Array in a Spread Spectrum Communications System," Proc. IEEE, Vol. 66, March 1978).
The primary advantage of demodulation-directed techniques lies in their efficient use of system resources and in their convergence speed. These algorithms rely on the demodulator-remodulator loop providing a very clean estimate of the desired signal. In most of these techniques, this requirement will be met after the demodulator has locked on to the received signal. However, until the demodulator does lock on, the reference signal estimate will generally be poor. For this reason, most demodulation-directed techniques are employed as tracking algorithms after a more sophisticated technique has been used to lock onto the desired signal. Many demodulation directed techniques encounter additional problems in dynamic environments where signals are appearing and disappearing over the course of the desired-signal transmission. In addition, these techniques are expensive to implement and are inflexible in their implementation, since they require a built-in demodulator matched to a specific desired signal to operate. This drawback renders them inapplicable to a system in which a variety of signals are of interest, e.g., in satellite transponders and reconnaissance systems.
The second class of techniques, referred to as channel-directed techniques, exploit known properties of the receiver channel or environment such as the spatial distribution of the received signals. In this class of techniques, knowledge of the receiver channel is used to generate and apply a reference signal to a conventional adaptation algorithm, or to estimate key statistics which are used to optimize the combiner weights. When applied to antenna arrays, most channel-directed methods exploit the discrete spatial distribution of the signals received by the array, i.e., the fact that the received signals impinge on the array from discrete directions of arrival. Examples of channel-directed techniques include the Griffiths P-vector algorithm (L. J. Griffiths, "A Simple Adaptive Algorithm for Real-Time Processing in Antenna Arrays," Proc. IEEE, Vol. 57, Oct. 1969) and the Frost constrained LMS algorithm (O. L. Frost, "An Algorithm for Linearly-Constrained Adaptive Array Processing," Proc. IEEE, Vol. 60, August 1972). Both of these techniques exploit the known direction-of-arrival of the desired signal. Techniques have also been devised to deal with the situation in which the direction-of-arrival of the desired-signal is unknown. Examples of such techniques are the generalized sidelobe canceller (I. J. Griffiths, M. J. Rude, "The P-Vector Algorithm: A Linearly Constrained Point of View," Proc. Twentieth Asilomar Conf. on Signals, Systems and Computers, Pacific Grovel, Calif., November 1986) and the signal subspace techniques referred to as MUSIC (R. O. Schmidt, "Multiple Emitter Location and Signal Parameter Estimation," Proc. RADC Spectrum Estimation Workshop, Oct. 1979. M. Wax, T. Shan, T. Kailath, "The Covariance Eigenstructure Approach to Detection and Estimation by Passive Arrays," IEEE Trans. ASSP, 1985) and ESPRIT (A. Paulraj, R. Roy, T. Kailath, "Estimation of Signal Parameters via Rotational Invariance Techniques--ESPRIT," Proc. Nineteenth Asilomnar Conf. on Signals, Systems and Computers, Pacific Grove, Calif., November 1985). These approaches can all be thought of as high-resolution spatial spectrum estimation techniques for locating lines in the received signal spatial spectrum.
All of the channel-directed techniques suffer from the common weakness that they require knowledge of the sensor geometry and/or the individual sensor or subarray characteristics to adapt the array. In practice, this characterization is usually obtained by a series of experiments, referred to as array calibration, to determine the so-called array manifold of the sensor network. The cost of array calibration can be quite high, and the measurement procedure is, in many applications, impractical. For example, a 16.times.16 planar array calibrated over a sphere with a one degree resolution in elevation and azimuth and using 16 bit accuracy requires approximately 128 megabytes of storage. This storage requirement increases exponentially as the number of search dimensions is increased, e.g., if the calibration is performed over temporal frequency or polarization in addition to elevation and azimuth. In addition, systems with considerably more sensors are desirable. Phased array communication systems have currently been proposed with over 10,000 elements, and current advances in low-cost microwave radio are pushing this figure higher. The storage requirements, not to mention the calibration times, for such large arrays renders these methods impractical. Furthermore, in certain applications, e.g., lightweight spaceborn arrays, airborne arrays, and towed acoustic arrays, the array geometry and even the sensor characteristics may be changing slowly with time; hence, an accurate set of calibration data may never be available.
Even when calibration data is available, the computational cost of using this data can be prohibitive. Both the generalized sidelobe canceller and MUSIC techniques require a search over the set of calibration data during the operation of the algorithm. In addition, the computational complexity of the MUSIC algorithm increases as the cube of the number of sensors in the array. For example, in the 16.times.16 sensor example cited above, the generalized sidelobe canceller algorithm would require 300.times.10.sup.6 real multiply-and-accumulate (RMAC) operations for a single exhaustive search, while the MUSIC algorithm would require 30.times.10.sup.9 RMACs for a single exhaustive search. Even if a 10 Mflop/sec (10 million floating point operations per second) computer were used, a computational time of 30 seconds would be needed using the generalized sidelobe canceller algorithm, and a time of one hour would be needed with the MUSIC algorithm. And, these times do not reflect the additional classification operations required to recognize the one desired signal among the multiple signals extracted by the algorithms.
The ESPRIT technique was pronosed in an effort to overcome these computational and storage problems. This technique does not require calibration data to operate and, hence, avoids many of the problems associated with the other channel-directed approaches. However, ESPRIT requires accurate knowledge of the noise covariance matrix of the received data, and its computational difficulty increases as the cube of the number of sensors in the array. For example, in the above example, the ESPRIT algorithm would require 20.times.10.sup.6 RMACs per search operation (about 2 seconds on a 10 Mflop/sec computer). The ESPRIT algorithm has a number of other shortcomings. It does not perform optimal signal extraction in the sense that the signal-to-noise-ratio is maximized, but instead, each output solution extracts a single signal and nulls the remaining coherent signals impinging on the array. The algorithm has difficulty extracting a single desired signal in a light interference background. In addition, the ESPRIT algorithm requires that the array elements be grouped into doublets with identical characteristics and common geometrical displacement. These conditions impose serious constraints on both the manufacture and performance of an ESPRIT array.
In the third class of techniques, referred to as set-theoretic property-mapping and property-restoral techniques, the output of the receiver is forced to possess a set of known properties possessed by the transmitted signal. Here, the receiver processor is adapted to restore know modulation properties of the desired signal to the processor output signal. Modulation properties are defined here as observable properties of the desired signal imparted by the modulation format used at the desired-signal transmitter. In many cases, these properties are destroyed by transmission over the communication channels. For example, the constant modulus property shared by FM, PSK, and CPFSK is destroyed by the addition of noise, other signals, or multipath interference to the transmitted signal. The property-restoral approach adapts a receiver processor to optimize an objective function that measures this property in the output signal.
The first use of the property-restoral concept was in Sato's algorithm (Y. Sato, "A Method of Self-Recovering Equalization for Multilevel Amplitude-Modulation Systems," IEEE Trans. Comm., vol. COM-23, pp. 679-682, June 1975) which was designed to equalize channel distortion in a BPSK telephony signal by minimizing the mean square error between the squared signal and unity. This algorithm was extended to restoral of general constant-modulus and QAM communication signals with the Constant-Modulus Algorithm (J. R. Treichler, B. G. Agee, "A New Approach to Multipath Correction of Constant Modulus Signals," IEEE Trans. ASSP, vol. ASSP-31, pp. 459-472, April 1983) and Godard's algorithm (D. N. Godard, "Self-Recovering Equalization and Carrier Tracking in Two-Dimensional Data Communication Systems," IEEE Trans. Comm., vol. COM-28, pp. 1867-1875, November 1980). More recently, set-theoretic property mapping (J. A. Cadzow, "Signal Enhancement--a Composite Property Mapping Algorithm," IEEE Trans ASSP, January 1988) has been advanced as a general technique for designing property-restoral algorithms.
The property-restoral algorithms described above have been successfully applied to adaptive signal extraction in both filters and antenna arrays and appear to have strong advantages over both the demodulation-directed and channel-directed techniques. However, these techniques still have drawbacks. The convergence and capture characteristics of all of these algorithms are still not well understood. In addition, the Constant-Modulus Algorithms are highly nondiscriminatory, requiring multitarget implementations to recover all the signals present in dense interference environments. The forrier drawback limits the application of these algorithms in automatic (unsupervised) communication systems where they must operate with a minimum of attention. The latter drawback is of critical importance in large-aperture systems and directed search applications, since it requires the algorithms to extract every signal in the environment using a multitarget implementation and to then classify the signals to find the one desired signal.
Broadly, it is an object of the present invention to provide an improved method for adapting receivers for the extraction of communications signals.
It is another object of the present invention to provide an apparatus and method for extracting signals that do not require a knowledge of the desired signal waveform.
It is yet another object of the present invention to provide an extraction apparatus and method that do not require a knowledge of the direction-of-arrival of the desired signal.
It is a still further object of the present invention to provide an extraction apparatus and method that do not require a knowledge of the background interference environment.
It is yet another object of the present invention to provide an extraction apparatus and method that do not require a knowledge of the geometry of the sensor array or of the individual sensor characteristics.
It is a still further object of the present invention to provide an extraction apparatus and method that can be programmed to sort and automatically extract signals with desired statistical properties from dense interference environments.
It is yet another object of the present invention to provide an extraction apparatus and method that require substantially less computation and storage than prior art competing methods.
It is a still further object of the present invention to provide an extraction apparatus and method that have unambiguous and well-understood convergence and capture properties.
These and other objects of the present invention will be apparent to those skilled in the art from the following detailed description of the invention and the accompanying drawings.